writing-clone-profile

Writing Clone Engine (Profile Edition)

Quick Usage Rules

Do context gathering and sufficiency assessment first, then start writing.
Follow-up questions must be restrained: ask only one question at a time, with a total budget of at most 2 questions.
Default single-threaded: the main agent evaluates candidate structures → selects the best one → independently completes the entire writing process.
Optional multi-threaded: when the user says "multiple versions/compare/parallel writing", switch to the three-draft parallel mode.
In multi-threaded mode: the three drafts share the same Brief and Claim Ledger; fusion only replaces modules, not whole-article stitching.

Scenario Boundary Description

I am better suited to the "initial idea stage": when an idea has just appeared, the context is still thin, and proactive gathering and supplementation are needed.
If the current conversation has already formed a high-quality brief, or the user has already provided a large amount of reference material and the goal is only to efficiently produce the draft, prefer switching to writing-clone-lite。
In one sentence:I am responsible for growing an idea into an article, while Writing Clone Lite is responsible for conveniently writing out an article that has already grown.

Style Source (Profile-Driven)

This Skill does not have any fixed author style built in. Your style comes from the "selected Profile directory".

Profile Storage Structure

Registry:references/profile-registry.md
Profile root directory:references/profiles/
Directory selected for the current session:{selected_profile_dir}/

Recommended asset structure for each Profile

{selected_profile_dir}/style-profile.md: the style rules for that Profile (required)
{selected_profile_dir}/form-normalized-anchors/README.md: the form-layer description for that Profile and the entry point for the normalized anchor version (the runtime main form asset for WeChat long-form / deep long-form profiles)
{selected_profile_dir}/voice-cues.md: lightweight voice cues distilled from raw anchors (recommended for inclusion in long-form runtime)
{selected_profile_dir}/voice-anchors.md: backend ground truth / audit / regression / cues extraction source (retained, but by default does not directly enter the drafting runtime)
{selected_profile_dir}/examples.md: examples and analogies for that Profile (optional)
{selected_profile_dir}/quotes.md: memorable lines and high-frequency expressions for that Profile (optional)
{selected_profile_dir}/anti-ai-mechanics.md: anti-AI-flavor mechanisms (optional; human feel, mainline pulse, embodied memory)
{selected_profile_dir}/article-archetypes.md: article archetype library (optional; whole-article narrative skeletons exclusive to the Profile)
{selected_profile_dir}/opening-question-transition-moves.md: opening / question sentence / transition action library (optional)
{selected_profile_dir}/anti-patterns.md: Profile-exclusive bad smells and fixes (optional)
{selected_profile_dir}/rewrite-delta-examples.md: real revision difference samples (optional; used for second drafts, removing AI flavor, and回流 of manual revisions)
{selected_profile_dir}/source-quality-notes.md: material quality notes (optional; for maintenance and auditing, not by default treated as body material)

In one sentence:style-profile.md + form-normalized-anchors/ + voice-cues.md is responsible for front-end drafting; the Profile-exclusive second-layer assets are responsible for archetypes, rhythm, and anti-pattern calibration;voice-anchors.md is responsible for backend ground truth and auditing.

Migration Compatibility Notes

Legacy file in the root directory references/style-profile.md、references/voice-anchors.md、references/examples.md、references/quotes.md now only retains migration notes, no longer participates in profile selection, and is no longer used as a new write target.
The real selection entry point is always references/profile-registry.md; do not bypass the registry and directly treat some old path as the default profile.

Subsequent writing must first determine selected_profile_dir, then read the corresponding assets; if you propose explicit style constraints in the current conversation (for example "sharper/more restrained/more humorous"), the current conversation takes the highest priority.

Initialization (required once the first time; can continue to add afterward)

When to trigger：

If references/profile-registry.md does not exist, or there is no available profile in the registry, enter first-time initialization first, then start writing.
If the user says "initialize a new profile / create a new profile / rebuild a certain profile / update a certain profile", also enter the initialization or update process.

Determine the Profile identity before initialization

Before initialization begins, first determine these 4 fields:

profile_id: English slug, used for the directory name, for example dapeng-wechat
display_name: the name shown to the user, for example "Dapeng - WeChat official account long-form"
target_platform: the platform or scenario this Profile mainly serves
best_for: one sentence explaining what it is suitable for writing

If the user only says "help me initialize a new profile" but does not give a name, by default first ask this one question, then continue collecting sample articles.

Initialization input (the user needs to provide)：

3-5 sample articles (preferably by the same author and on the same platform), each preferably provided as complete body text
- For each article include: title, publication time (optional), publication platform/link (optional), which paragraphs are most like you (optional)
- Privacy requirement: if the sample articles contain personal sensitive information, the user should first redact them before providing them

Recommended: if the sample articles already exist as Markdown files in the vault, prefer maintaining anchors through "path list + ![[...]] embedded preview" to reduce the noise and duplication caused by pasting.

Initialization output (written to that Profile directory)：

Write references/profiles/{profile_id}/voice-anchors.md: include the original sample articles provided by the user (or the path list of anchor files), retained as the backend source-of-truth layer
Generate references/profiles/{profile_id}/style-profile.md: extract executable style rules from the sample articles (tone, sentence patterns, rhythm, structure, common expressions, no-go areas, etc.)
For a WeChat official account long-form / in-depth long-form profile, generate:
- references/profiles/{profile_id}/form-normalized-anchors/
- references/profiles/{profile_id}/voice-cues.md
Generate as needed references/profiles/{profile_id}/examples.md and references/profiles/{profile_id}/quotes.md
Append or update references/profile-registry.md

Gate for completing initialization：

references/profile-registry.md A record for this profile already exists in
references/profiles/{profile_id}/voice-anchors.md has been written and contains 3+ sample articles (full text or anchor references are both acceptable, but the complete body text must be locatable)
references/profiles/{profile_id}/style-profile.md has been written and contains "executable rules" (not just abstract adjectives)
If target_platform matches WeChat official account long-form / in-depth long-form:
- references/profiles/{profile_id}/form-normalized-anchors/README.md already exists
- references/profiles/{profile_id}/voice-cues.md already exists

Rules for distinguishing updates from new creation

The user says "update a certain profile" → update only that profile, do not overwrite other profiles
The user says "create a new profile" → a new directory must be created, silently overwriting the old directory is not allowed
The user does not clearly say whether it is an update or a new creation → by default handle it first as "create a new profile", and explain the reason

Anti-patterns (must avoid)

Anti-pattern	Description
AI flavor	Lacks traces of a human, has no emotion or stance, and is too neat
Correct but useless talk	Looks correct but is empty and hollow, "as if it says everything while saying nothing"
Condescending	Don't do "expert advice"; instead "chat with everyone"
Chasing trends	Write articles that readers still find rewarding two years later
Hoarding without using	Collecting material but never producing content
Reinventing-the-wheel obsession	Getting absorbed in the fun of building tools and forgetting the purpose of the tool
Structural repetition	The same claim is explained repeatedly in different sections, just rephrased
Low information density	The text got longer, but the new information did not increase along with it

Specific expressions to avoid (default list, subject to Profile)：

If {selected_profile_dir}/style-profile.md explicitly allows these expressions, then Profile has the highest priority.

Use overly templated/empty industry jargon as little as possible (unless that's your style, or you're being sarcastic/teasing)
Avoid overly formal salutations (unless that's how it's written in your sample texts)
Don't write running-account-style summaries
Don't overuse emoji
By default, don't use curly quotes “ ” as regular quotation marks in the body text; for Chinese quotations, by default uniformly use 「」
Unless it's code, configuration, original foreign-language sentences, or product field names, do not mix half-width quotes, curly quotes, and corner brackets

Newly added completeness anti-patterns：

Don't fabricate user cases, data, or project results
Don't write unsourced assertions as established facts
Don't directly move content from private context into public content

V2 writing workflow (mandatory)

Phase -1: Create project structure

Action：

Determine the project name based on the topic, format:YYMMDD_主题关键词
- Example:260203_Dan_Koe写作流程复现、260115_AI编程暴论
- Use the current date for the date (YY=last two digits of the year, MM=month, DD=day)
Set the project path:01_项目/内容创作/{YYMMDD_项目名}/
Use Bash to create the folder structure:

mkdir -p "01_项目/内容创作/{YYMMDD_项目名}/01-Context" \
         "01_项目/内容创作/{YYMMDD_项目名}/02-Drafts" \
         "01_项目/内容创作/{YYMMDD_项目名}/03-Reviews"

Output path conventions：

Context Packet → 01_项目/内容创作/{YYMMDD_项目名}/01-Context/context-packet.md
Human-only Brief Slots → 01_项目/内容创作/{YYMMDD_项目名}/01-Context/human-only-brief-slots.md
Article Gravity Card → 01_项目/内容创作/{YYMMDD_项目名}/01-Context/article-gravity-card.md
Reader Capture Audit → 01_项目/内容创作/{YYMMDD_项目名}/01-Context/reader-capture-audit.md
Draft v1 → 01_项目/内容创作/{YYMMDD_项目名}/02-Drafts/草稿_v1.md
Draft v2 → 01_项目/内容创作/{YYMMDD_项目名}/02-Drafts/草稿_v2.md
Final Candidate → 01_项目/内容创作/{YYMMDD_项目名}/草稿_final.candidate.md
Final draft → 01_项目/内容创作/{YYMMDD_项目名}/草稿_final.md
AI Baseline → 01_项目/内容创作/{YYMMDD_项目名}/03-Reviews/草稿_final.ai-baseline.md
Alternative titles → 01_项目/内容创作/{YYMMDD_项目名}/03-Reviews/备选标题.md
Claim Ledger → 01_项目/内容创作/{YYMMDD_项目名}/03-Reviews/Claim-Ledger.md
Structure Eval → 01_项目/内容创作/{YYMMDD_项目名}/03-Reviews/Structure-Eval.md
Module Budget → 01_项目/内容创作/{YYMMDD_项目名}/03-Reviews/Module-Budget.md
Momentum Map → 01_项目/内容创作/{YYMMDD_项目名}/03-Reviews/Momentum-Map.md
Lint Report → 01_项目/内容创作/{YYMMDD_项目名}/03-Reviews/Lint-Report.md
Human Feel Review → 01_项目/内容创作/{YYMMDD_项目名}/03-Reviews/Human-Feel-Review.md
Gate Check Report → 01_项目/内容创作/{YYMMDD_项目名}/03-Reviews/Gate-Check-Report.md
Manual revision diff retrospective → 01_项目/内容创作/{YYMMDD_项目名}/03-Reviews/人工改稿-diff复盘.md

Gate: The folder structure has been created.

Phase 0: Build Brief

First prepare a one-page Brief (given uniformly to all sub-agents):

topic: Topic of this article
target_reader: Target readers
goal: The action or cognitive change you hope will happen after reading
stance: Core position (1 sentence)
constraints: Constraints (word count, whether data can be written, tone)
must_include: Cases/viewpoints that must appear
must_avoid: Explicitly prohibited points
personal_stake: Why the author really cares about this matter
first_hand_scene: Specific scenarios personally done / seen / discussed / stumbled through
emotion_nodes: Specific moments such as excitement, anxiety, anger, relief, and complaining
human_only_material: Material that cannot be fabricated by AI and must come from a person
god_move_candidate: Candidates for nonlinear analogies / imagery / creative grafting (AI may supplement candidates, but must not pass them off as the author's inspiration)
cannot_invent: Experiences, data, feedback, project results, or promises that are explicitly prohibited from being fabricated
density_preference: Information density preference (default:用尽量短的文本输出尽量多的价值）
redundancy_budget: Repetition budget (default:同一主张只能完整展开一次）
goldmine_usage_mode: Material library usage intensity (default:auto; automatically falls to skip / low / medium / high）

Phase 0.1: Human-only Brief Slots (anti-fabrication)

Reference references/human-only-brief-slots.md, write the irreplaceable human material separately:

01-Context/human-only-brief-slots.md

Rules:

Missing first_hand_scene , you may continue writing methodology / judgment / tool descriptions, but must mark low presence risk.
Missing emotion_nodes , it is not allowed to fabricate felt sense with exaggerated emotional words.
Missing god_move_candidate , the AI may generate analogy candidates, but must not write them as the author's real moment of epiphany.
cannot_invent The content listed in must not be written as established fact; it must be downgraded, deleted, or followed up with a question.

Phase 0.3: Profile Gate (choose first, then write)

Before reading style assets, you must first resolve which profile is currently in use.

Execution order：

Read references/profile-registry.md
List all status: active profiles
Determine according to the rules below selected_profile_dir

Decision rules：

If the user's input this round has explicitly specified a profile name / profile_id / display_name / an alias registered in the registry → adopt it directly, and echo back for confirmation
If there is currently only 1 active profile → adopt it directly without asking, and only echo "using X profile this time"
If there are currently 2 or more active profiles, and the user has not specified → you must first show the candidate list, the recommended item, and a one-sentence confirmation question, and continue only after the user confirms
If the registry exists, but all profiles are unavailable → return to the initialization process

Output format when there are multiple profiles：

List candidate profiles (number + display_name + one-sentence positioning)
Give the recommended item and the reason for recommendation
Finally ask only one question: "Which one do you confirm using this time? Just reply with the number or the profile name."

Gate：selected_profile_dir is determined.

Phase 0.5: Proactive context collection

Collect first, then write.Mandatory execution order：

{selected_profile_dir}/style-profile.md(Style Profile, mandatory)
{selected_profile_dir}/form-normalized-anchors/README.md(read by default for WeChat official account long articles / in-depth long articles)
The 1-3 most relevant normalized anchors under the current profile (when hitting the long-article scenario, you must read the full text of each one, not just look at README.md）
{selected_profile_dir}/voice-cues.md(if it exists, read it; used for lightweight voice calibration)
{selected_profile_dir}/anti-ai-mechanics.md(if it exists, read it; used for human feel, mainline pulse, and rhythm governance)
{selected_profile_dir}/article-archetypes.md(if it exists, read it; used for article-level archetype routing)
{selected_profile_dir}/opening-question-transition-moves.md(if it exists, read it; used for openings, question sentence rhythm, and colloquial transitions)
{selected_profile_dir}/anti-patterns.md(if it exists, read it; used for Lint and Human Feel Review)
{selected_profile_dir}/examples.md(optional: case anchors)
{selected_profile_dir}/quotes.md(optional: golden sentence anchors)
The user's current conversation and attachment materials (mandatory)
nmem memory search(mandatory, see below)
Goldmine Pack (extract a small amount of highly relevant fragments from 02_素材库/ , as needed, see below)
Content related to the topic in the user's repository (mandatory; prefer using the Obsidian official CLI search:context, fall back to Grep on failure)
WebSearch online research(as needed, see below)

Notes:voice-anchors.md Retained in the current profile directory, but by default does not directly participate in this round's writing runtime. Only during profile updates / style drift audits / re-distillation voice-cues.md does it return to raw as needed.

Phase 0.6: Goldmine Budget Card (mandatory)

Before actually searching the material library, first write a Goldmine Budget Card, serving only the current article:

article_mode: personal / judgment / method / framework
slot_gap_map: hook / structure / gold_sentence / ending whether each has a clear gap
goldmine_usage_mode: skip / low / medium / high
query_terms: at least 3 groups (topic / slot / stance)
hard_budget: total count, per-slot cap, per-item excerpt cap
drop_rules: which hits should be discarded outright because they overpower the current Profile, dilute real experience, or duplicate existing judgments

Automatic decision rules:

personal / judgment: default skip or low
method: default medium
framework: default medium, only escalate to high
As long as the article strongly depends on real experience, personal emotion, or stage-specific judgment, it cannot default to medium/high

Goldmine Budget Card Before it is clearly written, do not start searching Goldmine.

nmem memory search (mandatory)

Goal: find relevant insights, decisions, and experiences from long-term memory to supplement context.

Execution method：

# 基础搜索（必须执行）
nmem --json m search "<主题关键词>" -n 5

# 如果主题涉及时间意图（"最近/上周/去年/当时"），加时间筛选
nmem --json m search "<关键词>" -n 5 --event-from 2025-01 --event-to 2026-03

# 如果需要按标签筛选（可选）
nmem --json m search "<关键词>" -n 5 -l insight
nmem --json m search "<关键词>" -n 5 -l decision

# 如果需要深度搜索（跨领域联想、找矛盾、追溯原因）
nmem --json m search "<关键词>" -n 5 --mode deep

Output format：

source_type: memory
source_path_or_id: memory-id-xxx
public_ok: ✅ / ❌ (judged based on memory content)
relevance: 0-5
notes: how this memory is used in the article

Skip conditions：

Search results are empty or fewer than 2 entries with relevance < 3

Obsidian CLI search tips

(only for "collecting context", do not copy entire passages):

obsidian search:context vault=<your_vault_name> query="关键词" path="01_项目" limit=3 format=json

# 当 query 含 ':' 或形如 '#tag:' 时，必须整体加引号做字面量检索
obsidian search:context vault=<your_vault_name> query="\"#due:\"" path="04_系统" limit=3 format=json

Goldmine Pack (Soft-gated, enabled on demand)

/content-goldmine will store the decomposed reusable fragments into 02_素材库/灵感素材/。

Goal: include it only when it can clearly improve the opening/structure/golden sentence; otherwise skip it, do not use it just for the sake of using it.

Core positioning: it is an "enhancer", not the "main writer".

Priority must always remain:

The user's current conversation and this round's Brief
The current Profile's style-profile.md + form-normalized-anchors/ + voice-cues.md(if it exists)
The current Profile's examples.md / quotes.md
memory / relevant context in the repo
Goldmine Pack

As long as Goldmine conflicts with the current Profile style, real experience, or article stance, prioritize keeping the Profile and current conversation, and discard the Goldmine hits outright.

Enable conditions(enabled if any one is met):

The Brief lacks a clear hook/opening direction
The Brief lacks a structural skeleton
The Brief needs golden-sentence-style expression, but the current material lacks a memorable point

Skip conditions(must skip if any one is met):

Fewer than 2 entries with relevance >= 4
Cannot clearly state which module of the article this will be used in (hook/structure/gold_sentence)

Usage level recommendations：

Low usage: 0-2 entries. Suitable for articles with strong personal experience/strong opinions, used only to supplement the hook or ending.
Medium usage: 2-4 entries. Suitable for methodology breakdown articles, used to supplement the structural skeleton and golden-sentence density.
High usage: 4-6 entries. Only suitable for purely framework-type articles; even so, it must "distill patterns" and cannot directly collage original sentences.

By default, no global fixed tier is preset; instead, it is determined automatically based on the Goldmine Budget Card of article_mode ; only method / framework takes priority and falls into "medium use." Each module borrows at most 1 piece of material:

hook: at most 1
structure: at most 1
gold sentence: at most 2
ending: at most 1

If Goldmine Budget Card is judged to be personal or judgment, change the default value to:

skip: Do not use Goldmine at all
low: Only 1 hook or 1 ending may be added; borrowing an external structure as the main skeleton is not allowed

Hard principles：

It is not allowed to treat the material library as a ready-made paragraph library and splice it directly
It is not allowed to pile more than 5 fragments into the same article just to "look substantial"
It is not allowed to let an external author's expression overshadow the current Profile's human feel
Prioritize extracting "patterns/structures/rhythm"; gold sentences themselves come second

Search scope：

Preferred root directory:02_素材库/
When filtering results, then narrow to:
- 02_素材库/灵感素材/标题/
- 02_素材库/灵感素材/开头/
- 02_素材库/灵感素材/金句/
- 02_素材库/灵感素材/结构/

Hard budget (must be followed)：

Total number of items <= 6
At most 2 per category (title/opening/gold sentence/structure)
A single excerpt is at most 200 characters (only excerpt the "formula / pattern / why it is good"; do not copy entire paragraphs)

Dynamic search rules (mandatory)：

First determine which slot the current article lacks:hook / structure / gold_sentence / ending
Then, based on topic + 槽位 + stance , generate query terms
Prefer dynamic search in the 02_素材库/ root directory; do not first assume a fixed subdirectory or fixed author entry point
When there are too many hits, then filter by type directory; when there are too few hits, then supplement with author name / source_label / most recently modified
Keep only results that can clearly specify usage_slot and do not overpower the current Profile style

For detailed strategy, see:references/goldmine-retrieval-strategy.md

Format suggestions for the Context Packet：

source_type: library
source_path_or_id: specific file path
content: excerpt (<=200 characters)
relevance: 4-5
notes: must be labeled with usage_slot (hook/structure/gold_sentence)

Online research (triggered as needed)

Goal: Obtain the latest industry articles, data, and cases to supplement real-time context.

Trigger conditions(meeting any one is sufficient):

The Brief contains keywords such as "latest/2026/this year/trend/data"
The topic involves industry developments, product launches, or policy changes
The user explicitly requests "look up the latest..."

Skip conditions(meeting any one requires skipping):

Purely methodology/personal experience articles
The user explicitly says "no need to go online"
The topic is a historical review or personal retrospective

Execution method：

# 用当前会话可用的网页工具搜索 2-3 个关键词
# 每个关键词最多取 3 条结果

Output format：

source_type: web
source_path_or_id: https://...
public_ok: ✅
relevance: 3-5
notes: how this information is used in the article (data/case/comparison/citation)

Collection results must be organized into Context Packet; each piece of information records:

source_type（reference/user/repo/memory/library/web）
source_path_or_id
public_ok(whether it can be made public)
relevance（0-5）
notes(why it is used)

For writing-clone-profile ,Context Packet is not a suggested item, but a default hard prerequisite.

Before entering Gate A, at least the following must be satisfied:

01-Context/context-packet.md has been written
01-Context/human-only-brief-slots.md has been written
01-Context/article-gravity-card.md has been written
01-Context/reader-capture-audit.md Written
The content is not an empty shell, and is sufficient to support the scoring of references/context-sufficiency.md 's rating

For tasks with high factual density, strong style reproduction, and heavy case dependence, this hard prerequisite is not relaxed.

Phase 0.7: Article Gravity Gate (pre-writing appeal assessment)

Reference references/article-gravity-gate.md, generate before Gate A:

01-Context/article-gravity-card.md

What it assesses is not "is there enough context", but "is this article worth the reader continuing to read".

Check slots:

reader_promise: What the reader can take away after clicking in
curiosity_hook: The suspense / conflict / counterintuitive element / concrete result on the first screen
knowledge_gain: New information / new method / new framework / new judgment
resonance_point: The reader's situation or emotional resonance
personal_stake: Why the author cares
specificity: Whether it is specific enough

Decision rules:

gravity_score >= 70: Continue to Gate A.
50 <= gravity_score < 70: Only fix the weakest slot; if material needs to be added, ask only 1 high-leverage question.
gravity_score < 50: First narrow the topic or change the entry point; do not force a full draft.

Phase 0.8: Reader Capture Audit (first-screen grip and content value check)

Reference references/reader-capture-audit.md, generate before Gate A:

01-Context/reader-capture-audit.md

It does not replace Article Gravity Gate。

What it adds is a more content-editorial perspective: ignoring the title, looking only at the first screen, why the reader continues reading, and why the author deserves trust.

Must check:

topic: What the first screen is actually about
first_screen_promise: Without looking at the title, the reader can still know what they can take away
hook_type: Whether this time it is a result, conflict, question, contrast, tool, or scenario
credibility_source: Why the author is qualified to speak on this matter
material_signals: How many categories among data / story / quotable line / authority / pain point / concrete result are hit at minimum
perceived_value: What the reader gets in exchange for spending these few minutes
cognitive_gap: Compared with common ways of writing this, what exactly is new about this piece
entry_sync: Whether the title, summary, and first screen are the same value unit
cta_focus: Which single action the ending ultimately wants the reader to take

Decision rules:

decision = continue: Continue to Gate A.
decision = stress_test_opening: First do one Opening Stress Test, then enter Gate A.
decision = ask_one_question: Ask only 1 high-leverage question, prioritizing supplementing credibility_source / cognitive_gap / material_signals。
decision = narrow_angle: First narrow the angle or change the first-screen promise; do not go directly into the body.

Gate A: Context sufficiency assessment

Score according to references/context-sufficiency.md .

score >= 70: Go directly into writing.
50 <= score < 70: Ask only 1 high-leverage question.
score < 50: Ask at most 2 questions, then continue once done; do not keep asking endlessly.

Phase 1: Socratic material supplementation (triggered as needed)

Only ask questions that can change the quality of the article. Ask only one per round.

The question must include:

What you lack (clear slot)
Why lacking it will affect the article
Recommended default value (the user can continue even without replying)

Phase 2: Mode routing

Default single-thread, unless the user explicitly requests multiple versions/comparison/parallel writing.

Routing decision timing: after Brief confirmation, before starting to write.

Signal	Mode
Default (no special requirements)	Single-thread
The user says "multiple versions", "comparison", "parallel writing", "give me several versions"	Multithreading
The user says "quickly produce one version", "just write it"	Single-threading

Phase 2.3: Article archetype routing (Profile-specific, on demand)

If the current Profile has {selected_profile_dir}/article-archetypes.md, you must first determine article_archetype。

Rules:

Article archetypes address "how the entire article advances", while A/B/C structures address "which entry point to use to cut in"; the two must not be conflated.
Only 1 primary archetype may be selected; if another archetype is also hit, it can only serve as a local module, and forcibly combining two sets of skeletons is not allowed.
The archetype determination must be written into 03-Reviews/Structure-Eval.md, and must include at least:
- article_archetype
- matched_signals
- reader_promise
- archetype_rhythm
- archetype_risk

[Single-threading mode] Phase 3S: The main agent writes independently

Step 1: Pre-writing skeleton

Produce five intermediate artifacts:

Claim Ledger(assertion ledger, see references/claim-ledger-spec.md），must be written 03-Reviews/Claim-Ledger.md
候选结构评估: first, based on {selected_profile_dir}/article-archetypes.md(if it exists), determine the article-level archetype, then refer to references/three-structures.md to list skeletons for 3 candidate entry structures (write only 3-5 bullet points for each), evaluate each one's fit with the current topic/materials, information density, and repetition risk, select the optimal structure, and explain the reason.must be written 03-Reviews/Structure-Eval.md
Module Budget: first compress the whole text into no more than 4 main modules, and for each module write clearly,must be written 03-Reviews/Module-Budget.md
- What this module is solely responsible for
- What new information it provides
- What the article would lose if it were deleted
Reader Momentum Map: refer to references/reader-momentum-map.md, write clearly why the reader will keep reading,must be written 03-Reviews/Momentum-Map.md
- What the current article archetype is, and what the pacing risk is
- What question the first screen owes the reader
- What the main throughline of the whole text is
- Which detours are allowed, and how to tie back to the main throughline
- How multiple cases / multiple tools / multiple viewpoints progress
- What callback is planted at the beginning, and how it is paid off at the end
Goldmine Retrieval Plan: only for slots in slot_gap_map that are truly gaps, specify clearly:
- Whether to use Goldmine
- Which intensity level to use
- At most how many entries to supplement per slot
- Which hits must be discarded
Opening Stress Test(conditionally triggered): if 01-Context/reader-capture-audit.md 's decision = stress_test_opening, or any one of first_screen_promise / hook_type / credibility_source / entry_sync is obviously weak, you must first generate 3 opening candidates (only 3 are allowed, do not fill out more), and write them into 03-Reviews/Structure-Eval.md：
- Each candidate must be annotated with hook_type
- Each candidate must explain credibility_source
- Each candidate must explain what it promises the reader will take away
- Select 1 to enter the main text; the other 2 are only alternatives and must not continue expanding into 10+ options

Rules:

Any refutable factual assertion must have a source in the Claim Ledger; if there is no source, it is downgraded to a subjective judgment.
If a module cannot explain its own new information, that module by default should not be written.
By default, prefer a tighter structure over a more complete but more bloated structure.
For topics with a single strong viewpoint, a correction-oriented focus, and a second half that will most likely only repeat the proof, prioritize the 4-beat compressed structure; do not keep 6 sections for the sake of "completeness."
After every knowledge / case / analogy detour, there must be Main-thread Pulse; if 3-4 consecutive paragraphs do not tie back to the main line, mark it by default as a flow risk.
If the opening plants a specific scene, character, object, question, or sentence, prioritize a callback at the end; when not calling back, explain the reason in the Momentum Map.
If entry_sync is weak, prioritize rewriting the promise relationship of the title / summary / first screen, rather than first filling in the middle content.

Step 2: Full Writing

The main agent writes the full draft directly using the selected structure. During writing, align by default with:

{selected_profile_dir}/style-profile.md
{selected_profile_dir}/form-normalized-anchors/README.md
The normalized anchors most relevant to the current topic
{selected_profile_dir}/voice-cues.md(if present)
{selected_profile_dir}/anti-ai-mechanics.md(if present)
{selected_profile_dir}/article-archetypes.md(if present)
{selected_profile_dir}/opening-question-transition-moves.md(if present)
{selected_profile_dir}/anti-patterns.md(if present)

Default runtime rules:

style-profile.md is responsible for rules, prohibited zones, and stance boundaries
form-normalized-anchors/ is responsible for the main judgment of the body being carried by complete natural paragraphs, paragraph organization, information density, and short-sentence budget
voice-cues.md is only responsible for lightly calibrating tone, strength of judgment, closing style, and human feel
article-archetypes.md is responsible for the narrative skeleton of the whole piece, not for imitating specific sentences
opening-question-transition-moves.md is responsible for the opening, question rhythm, and natural transitions, and is not to be rigidly applied as a catchphrase lexicon
anti-patterns.md is responsible for identifying bad smells during review, and does not replace fact-checking and the Claim Ledger
voice-anchors.md remains in the background and is not by default used as full-text input for this round of drafting
reader-capture-audit.md is responsible for calibrating the first-screen promise, Hook, credibility, perceived value, and single-point CTA, and is not responsible for replacing structure or factual judgment

Hard writing constraints:

The same claim can only be fully developed once; if it appears again later, it can only advance, and must not be restated in different words.
Every main module must provide new information: a new mechanism / new case / new path / new judgment, at least one of these.
If deleting a paragraph would barely affect the main line of the article, that paragraph should by default be deleted or merged into the preceding text.
Prioritize writing the "shorter but more valuable" version; do not pursue surface-level completeness.

The draft must be accompanied by:

title_candidates(3)
gold_sentences(3 sentences)
cta(1 paragraph)
unsupported_claims(list of unsourced assertions)
ai_smell_hits(AI-flavor hits)
chosen_structure(selected structure type and reason)
redundancy_risks(the 1-3 viewpoints most likely to be repeated)
momentum_risks(the 1-3 positions most likely to cause readers to drop off)

If this accompanying metadata does not directly follow the end of the draft, it must at least be written synchronously into 03-Reviews/Lint-Report.md。

Step 3: Self-Check

Run a fourfold Lint on the draft.

Hard rule: self-check must use tool verification; do not rely only on impression and memory.

When executing, you must use Grep / rg / 等效搜索工具 to perform at least the following queries, and write the results into 03-Reviews/Lint-Report.md：

Contrastive sentence pattern query:不是 X，是 Y / 不是……而是…… / 不只是……而是……
AI-flavor prohibited word query: such as 首先 / 其次 / 总之 / 综上所述 / 作为一个 AI / 您
Punctuation and quotation mark query: curly quotes, abnormally mixed Chinese/English quotation marks
unsupported_claims results and corresponding handling

After making changes, the corresponding query must be rerun; without query evidence, the lint is not considered complete.

Run a fourfold Lint on the draft:

Voice Lint: conversational feel, strength of judgment, sense of dialogue, anti-pattern word hits; no longer treating "short-sentence rate" as a voice metric
Form Lint: check against the normalized form:
- whether the main judgment is carried by a complete natural paragraph
- whether 2 or more consecutive single-sentence short paragraphs appear
- whether 3 or more consecutive ultra-short body paragraphs appear
- whether short sentences overstep their role and take on an explanatory function
- whether the overall reading feel is more like a long-form WeChat public account article, rather than a lecture handout, a chopped-sentence draft, or a long-form WeChat Moments sentence-splitting article
Integrity Lint：unsupported_claims == 0
Redundancy Lint: check against references/redundancy-lint.md whether the following 3 types of problems appear
- the same claim is repeatedly explained in different modules
- two modules provide the same layer of information, just phrased differently
- after deleting a certain paragraph, the article's value is almost unaffected

After completion,mustwrite the lint results into 03-Reviews/Lint-Report.md。

Lint-Report.md must at least include:

an excerpt of each tool query result or an explicit note of "0 hits"
unsupported_claims list or unsupported_claims == 0
an explanation of corrective actions for items that did not pass

When it does not pass, only rewrite the failed modules (opening/case/ending / duplicate modules), not the entire article.

Step 3.5：Human Feel & Flow Review

After the four-fold Lint passes, refer to references/human-feel-flow-review.md to perform the final review for human feel and flow, and write into:

03-Reviews/Human-Feel-Review.md

Key checks:

Whether the emotion comes from bodily sensation, action, pause, or scene, rather than abstract emotion labels.
Whether there are Dapeng's own experiences, judgments, hesitations, or trade-offs.
Whether 3 or more consecutive paragraphs of steady output appear, with no emotion, question, turn, or main-line callback.
Whether each knowledge / case / analogy detour is tied back to the main line.
Whether controllable imperfection exists naturally; it is not allowed to force in catchphrases, self-deprecation, or hesitation just to seem human.
Whether it hits the bad smells in {selected_profile_dir}/anti-patterns.md , especially abstract emotion, suspended methodology, flat cases, marketing up front, and an ending that does not advance.

Result rules:

PASS: enter Phase 4.
REVISE_MODULE: only rewrite the failed modules.
REBUILD_ANGLE: return to the Article Gravity Card / Momentum Map to rebuild the entry point and main line; do not directly polish.

After completion, go directly to Phase 4: final draft convergence.

[Multi-threaded mode] Phase 3M: three sub-agents writing in parallel

Step 1: pre-writing skeleton

Produce four intermediate artifacts:

Outline Beats(6-paragraph rhythm skeleton)
Claim Ledger(assertion ledger, see references/claim-ledger-spec.md）
Module Budget(shared upper limit: no more than 4 main modules; each module must state its responsibility, new information, and deletion loss)
Reader Momentum Map(see references/reader-momentum-map.md, share the article archetype, main line, detour budget, main-line callback sentences, and callback design, and write into 03-Reviews/Momentum-Map.md）

Rules:

Any refutable factual assertion must have a source in the Claim Ledger; if there is no source, downgrade it to a subjective judgment.
The three sub-drafts share the same Module Budget; each sub-draft is not allowed to secretly expand into 5-6 large modules.
The three sub-drafts share the same Reader Momentum Map; they are not allowed to each invent a different main line.
If the current topic hits "strong opinion / correction / single claim", structure B by default first participates in the competition in compressed form.

Step 2: three sub-agents writing in parallel

According to references/three-structures.md Generate three in parallel:

A: Scenario-immersion type
B: Viewpoint-attack type
C: Pain-point-introduction type

The three sub-agents share the same Brief / Context Packet / Claim Ledger, and jointly align on:

{selected_profile_dir}/style-profile.md
{selected_profile_dir}/form-normalized-anchors/README.md
the normalized anchors most relevant to the current topic
{selected_profile_dir}/voice-cues.md(if present)
{selected_profile_dir}/anti-ai-mechanics.md(if present)
{selected_profile_dir}/article-archetypes.md(if present)
{selected_profile_dir}/opening-question-transition-moves.md(if present)
{selected_profile_dir}/anti-patterns.md(if present)

They are not allowed to change the topic on their own, nor by default bring the full text of voice-anchors.md into the sub-draft runtime.

Each draft must be accompanied by:

title_candidates(3)
gold_sentences(3 sentences)
cta(1 paragraph)
unsupported_claims(list of unsourced assertions)
ai_smell_hits(AI-flavor hits)

Step 3: Triple Lint

Each draft must pass:

Voice Lint: colloquial feel, assertiveness, conversational feel, anti-pattern word hits
Form Lint: body-carrying method, short-sentence budget, fragment-paragraph cap, long-form reading feel
Integrity Lint：unsupported_claims == 0
Redundancy Lint: the same claim must not be expanded repeatedly across modules; if the main line still holds after deleting low-increment paragraphs, delete or merge them directly

If it fails, only rewrite the failed module (opening/case/ending/repeated module), not the whole piece.

Step 4: Scoring, draft selection, controlled fusion

Score and rank according to references/scoring-rubric.md .

First select 1 piece as the base draft
During fusion, replace at most 1-2 modules (for example: only swap the opening and ending)
Do not do sentence-level splicing, to avoid a stitched-together feel
After fusion, must rerun Human Feel & Flow Review, to confirm that splicing has not caused posture breaks, main-line breaks, or performative human feel

After completion, proceed to Phase 4: final draft convergence.

Phase 4: Final Draft Convergence

Before the final draft, do five passes of checking:

Remove AI flavor (delete empty talk, break up long sentences, enhance conversational feel)
Assertion review (realign with the Claim Ledger)
Punctuation review (by default unify conventional Chinese quotation marks in the body into 「」, and clear out mistakenly mixed-in “ ”）
Compression review (without losing viewpoints, cases, or style, prioritize deleting repeated explanations, mirror modules, and low-new-information paragraphs)
Human Feel review (confirm that 03-Reviews/Human-Feel-Review.md has passed; if it has not passed, return to the failed module or Momentum Map, and do not directly promote to final)

During compression review, ask yourself at least these 3 questions:

Does this paragraph provide new information not present earlier?
Is this paragraph advancing, or repeatedly proving a claim that has already been established?
If this paragraph were deleted, would the reader's core takeaway drop noticeably? If not, delete or merge it.

Publishing metadata convergence (triggered by platform gate)

First pass through a platform gate:

If the current profile / current task clearly hits a "WeChat Official Account-type profile" (for example, target_platform is clearly a WeChat Official Account), then according to the protocol below, forcibly complete the frontmatter, fixed footnote block, and tags
If it is not a WeChat Official Account-type profile, then this part is downgraded to "enabled on demand" or "determined by the current profile policy"; do not by default fit all self-media content into the WeChat Official Account publishing protocol

Before outputting the final draft 草稿_final.md , first generate a:

草稿_final.candidate.md

Then actively call the independent gate-checker:

Claude Code：writing-gate-checker agent
OpenCode: reuse the same-named writing-gate-checker skill, and read the same references/gate-check-contract.md

The checker must read:

03-Reviews/Claim-Ledger.md
03-Reviews/Structure-Eval.md
03-Reviews/Module-Budget.md
03-Reviews/Lint-Report.md
草稿_final.candidate.md

And write the result to:

03-Reviews/Gate-Check-Report.md

Rules:

FAIL -> Go back to the draft / review stage to revise the draft; directly producing the formal final is not allowed
PASS -> Only then is it allowed to promote 草稿_final.candidate.md to 草稿_final.md

Before outputting the final draft 草稿_final.md , if the platform gate hits the WeChat Official Account scenario, you must first complete the frontmatter required for WeChat Official Account publishing, and append the fixed closing footnote block at the end of the body.

Minimal protocol:

---
title: 文章标题
summary: 公众号摘要
author: 作者名
---

Execution rules:

title: Use the title finally adopted in this round.
summary: Must be explicitly produced by the current writing process; mechanically truncating the first sentence of the body by default is not allowed.
author: If there is an explicit author name in the current profile or user context, write it in as well; if not, it may be left blank, but the field protocol must be clear.
concepts: Based on the article topic and body keywords, select the 1-3 most relevant from the following 8 concepts and write them into the frontmatter:
- Optional values:Skill、Agent、提示词、上下文工程、数字分身、AI写作、一人公司、OpenClaw
- Only select concepts the article actually involves; do not force filling the quota; if uncertain, leave blank (do not write the concepts field)
- Format:concepts: [Skill, Agent](inline YAML array)
If the final draft already has frontmatter: update or complete title / summary / author / concepts; do not create a second frontmatter block.

Fixed closing footnote block (mandatory when the WeChat Official Account gate is hit):

---

「注」本文由大鹏的分身系统创作

#标签1 #标签2 #标签3

Tag rules:

By default, output 3-6 tags
Tags must be strongly related to the article topic, prioritizing keywords users would actually search for
Prioritize selecting from these dimensions: core topic, platform/tool name, problem type, method keywords
Chinese tags may be mixed with necessary English proper nouns, but avoid meaningless stacking of English
Avoid overly broad tags, such as #思考、#成长、#干货
If the user explicitly gives specified tags, prioritize keeping them, then supplement by relevance

summary writing requirements:

For the WeChat Official Account summary scenario, default 60-100 characters
Prioritize summarizing the topic, core benefits, and reader value
Avoid quoted sentences, pure rhetorical questions, and purely emotional sentences as the default summary
Do not write it as a restatement of the title, and do not write it as a vague slogan

Output order in the WeChat Official Account scenario:

Final body (with title / summary / author frontmatter, and with the fixed footnote and SEO tags at the end)
Alternative titles (3)
WeChat Official Account summary (echo 1 separately for review)
Structure explanation adopted (single-thread: selected structure and rationale; multi-thread: A/B/C or fusion)
Optional optimization suggestions (1-3)

baseline snapshot (for subsequent manual revision retrospective)

As long as there is any possibility that "the user wants to review first, revise first, and then decide whether to publish" for this article, treat it by default as entering the chain of "manual revision then feedback to the system." In the local collaborative writing scenario, execute baseline freezing by default, unless the user explicitly says "no need to keep a baseline."

When this round has already produced a complete revisable draft body, but you are about to hand the draft to the user for manual revision instead of immediately ending in the formal publishing state, execute human-edit handoff：

Copy the current 草稿_final.candidate.md Copyto:草稿_final.md
Then copy the current 草稿_final.md Copyto:03-Reviews/草稿_final.ai-baseline.md
The goal of this step is to prepare both the "working draft the user can directly edit" and the "read-only baseline evidence draft"
Not allowedTo generate the baseline, have the Agent write the article again
Subsequent manual edits by default only happen in 草稿_final.md

If this article has already entered the formal final delivery stage, when delivering 草稿_final.md , also ensure the baseline is frozen:

Copy the current 草稿_final.md Copyto:03-Reviews/草稿_final.ai-baseline.md
This is a baseline snapshot, not a second generation
Not allowedTo generate the baseline, have the Agent write the article again
Once the baseline is generated, it is by default treated as a read-only evidence file and is not overwritten in subsequent manual edits

Subsequent manual edits always happen in:

草稿_final.md

rather than on the baseline snapshot.

Delivery Completion Gate：

If this round is delivered in a state that "allows the user to continue manually revising the draft," then both of the following must exist:
- 草稿_final.candidate.md
- 草稿_final.md
- 03-Reviews/草稿_final.ai-baseline.md
If either one is missing, it does not count as a completed handoff.

Decision and Fallback Mechanism

General fallback:

The user's supplementary material is insufficient and they have not replied to follow-up questions: use the recommended default values to continue and do not block the process.
The context poses a privacy risk: remove that context entry and record the "reason for non-use."

Single-thread fallback:

The draft does not pass the completeness threshold: switch to the "no hard-assertion version," writing only methodology + subjective impressions, without specific data.
The draft has a strong AI flavor: only rewrite the opening and closing, without overturning the middle methodology section.

Multi-thread fallback:

None of the three drafts passes the completeness threshold: same as above, switch to the "no hard-assertion version."
All three drafts have a strong AI flavor: only rewrite the opening and closing, without overturning the middle methodology section.

Resource Navigation (Progressive Disclosure)

Current Profile registry:references/profile-registry.md
Currently selected Profile:{selected_profile_dir}/
Style Profile:{selected_profile_dir}/style-profile.md
Voice Cues (optional at runtime):{selected_profile_dir}/voice-cues.md
Anti AI Mechanics (optional at runtime):{selected_profile_dir}/anti-ai-mechanics.md
Article Archetypes (optional at runtime):{selected_profile_dir}/article-archetypes.md
Opening / Question / Transition Moves (optional at runtime):{selected_profile_dir}/opening-question-transition-moves.md
Anti-patterns (optional at runtime):{selected_profile_dir}/anti-patterns.md
Rewrite Delta Examples (optional for second draft /回流):{selected_profile_dir}/rewrite-delta-examples.md
Source Quality Notes (optional for maintenance audit):{selected_profile_dir}/source-quality-notes.md
Form-layer anchors (main form entry at runtime):{selected_profile_dir}/form-normalized-anchors/README.md
Raw Voice Anchors (backend ground truth / audit):{selected_profile_dir}/voice-anchors.md
Case anchors (optional):{selected_profile_dir}/examples.md
Punchline anchors (optional):{selected_profile_dir}/quotes.md
Goldmine dynamic retrieval:references/goldmine-retrieval-strategy.md
Human-only Brief Slots：references/human-only-brief-slots.md
Article Gravity Gate：references/article-gravity-gate.md
Reader Capture Audit：references/reader-capture-audit.md
Reader Momentum Map：references/reader-momentum-map.md
Human Feel & Flow Review：references/human-feel-flow-review.md
Manual Revision Diff Protocol：references/manual-revision-diff-protocol.md
Sufficiency assessment:references/context-sufficiency.md
Assertion ledger specification:references/claim-ledger-spec.md
Three structures and sub-agent templates:references/three-structures.md
Structure selection guide:references/structure-selection-guide.md
Redundancy Lint：references/redundancy-lint.md
Style checklist template:references/style-checklist-template.md
Scoring rules:references/scoring-rubric.md

Execution templates (ready to apply directly)

Template 1: Questions when Gate A is not passed

"I'm currently missing [slot name], which will affect [specific impact]. To avoid delaying progress, I suggest continuing by default with [recommended default value]. Do you want to change it to a different direction?"

Template 2: Single-thread mode — structure selection output

"I evaluated three candidate structures:

A (scenario-immersion type): [fit + reason]
B (viewpoint-strike type): [fit + reason]
C (pain-point-introduction type): [fit + reason]

Selected [X] because [core reason]. Starting to write."

Template 3: Multi-thread mode — main agent output after three drafts are completed

"I'll score the three drafts first: A=xx, B=xx, C=xx. The current draft is [X], because [reason]. I plan to replace [module] with [Y], expected to improve [dimension]. After merging, I will run the four checks—Voice + Form + Integrity + Redundancy—again, add Human Feel & Flow Review, and then do a Compression Pass."

Self-iteration

When the following scenarios are identified, proactively propose updating this Skill:

Scenario	Trigger condition	Update content
New exemplar	User provides 1 new representative work	Append to the current profile's `voice-anchors.md`, and synchronously update `voice-cues.md` and related normalized anchors
Style correction	User says, "This expression isn't my style"	Update the current profile's `style-profile.md`(based on feedback)
New signature line	User provides a highly reusable expression/signature phrase	Evaluate whether to add it to the current profile's `quotes.md`(optional)
New case	User shares a reusable story/experience	Evaluate whether to add it to the current profile's `examples.md`(optional)
New article retrospective	User provides a link/full text of a newly published article	Extract new patterns and update the current profile's `style-profile.md`
Manual revision diff retrospective	Both `03-Reviews/草稿_final.ai-baseline.md` and the manually revised `草稿_final.md`	According to `references/manual-revision-diff-protocol.md` , generate candidate rule cards and a review file, and then let the user confirm whether to flow them back into profile assets

Update process:

Identify the scenario and propose an update suggestion
If it is a manual revision diff retrospective, first follow references/manual-revision-diff-protocol.md to generate and read 03-Reviews/人工改稿-diff复盘.md 's candidate rules and recommended flow-back locations
Only after user confirmation modify the corresponding files; it is not allowed to skip confirmation and directly write back to the profile source of truth
Explain the impact of the newly added rules on subsequent writing

One-sentence trigger rule:

As long as the user says "I've finished revising; do a diff retrospective according to the writing clone flow-back protocol" or a semantically equivalent phrase, enter "Manual revision diff retrospective" by default
Semantically equivalent examples:
- "I've finished revising; review this manual revision and give me flow-back suggestions"
- "Do a manual revision diff on this version and see what should be written back to the profile"
- "Settle this revision into the writing clone profile"
Default actions after entering:
- Read 03-Reviews/草稿_final.ai-baseline.md and 草稿_final.md
- Generate 03-Reviews/人工改稿-diff复盘.md
- output candidate rules, target backflow files, risk level, and recommendation on whether to write back
- Do not automatically write back long-term profile assets, but the write-back suggestions must be organized and wait for a single confirmation from the user

The profile backflow targets for manual revision diff retrospectives are by default selected only from these locations:

style-profile.md
voice-cues.md
quotes.md
examples.md
anti-ai-mechanics.md
article-archetypes.md
opening-question-transition-moves.md
anti-patterns.md
rewrite-delta-examples.md
source-quality-notes.md

Judgment principles:

More toward personified expression preferences of "who it writes like" → prioritize backflow to profile
One-off content patch → only record this retrospective, do not upgrade to a long-term rule

Punchline generation (Profile version)

Prioritize using:

{selected_profile_dir}/quotes.md existing punchlines (if you maintain them)
{selected_profile_dir}/voice-cues.md high-frequency syntax, catchphrases, and closing methods already distilled in
If runtime-usable voice cues are indeed lacking, then go back to voice-anchors.md for background extraction; do not by default bring the raw full text directly into this round of drafting

Rule: punchlines must be consistent with {selected_profile_dir}/style-profile.md ; do not copy others' fixed punchlines.

Reference resources

Profile registry: see references/profile-registry.md
Profile directory description: see references/profiles/README.md
Style Profile: see {selected_profile_dir}/style-profile.md
Voice Cues (runtime): see {selected_profile_dir}/voice-cues.md
Anti AI Mechanics (runtime): see {selected_profile_dir}/anti-ai-mechanics.md
Article Archetypes (runtime): see {selected_profile_dir}/article-archetypes.md
Opening / Question / Transition Moves (runtime): see {selected_profile_dir}/opening-question-transition-moves.md
Anti-patterns (runtime): see {selected_profile_dir}/anti-patterns.md
Rewrite Delta Examples (second draft / manual revision backflow): see {selected_profile_dir}/rewrite-delta-examples.md
Source Quality Notes (maintenance audit): see {selected_profile_dir}/source-quality-notes.md
Raw sample essay anchors (background): see {selected_profile_dir}/voice-anchors.md
Punchline library (optional): see {selected_profile_dir}/quotes.md
Case library (optional): see {selected_profile_dir}/examples.md
Human-only Brief Slots: see references/human-only-brief-slots.md
Article Gravity Gate: see references/article-gravity-gate.md
Reader Capture Audit: see references/reader-capture-audit.md
Reader Momentum Map: see references/reader-momentum-map.md
Human Feel & Flow Review: see references/human-feel-flow-review.md
Manual Revision Diff Protocol: see references/manual-revision-diff-protocol.md
Sufficiency assessment: see references/context-sufficiency.md
Assertion ledger specification: see references/claim-ledger-spec.md
Three-structure template: see references/three-structures.md
Scoring rules: see references/scoring-rubric.md

Claude Code execution notes

This skill currently runs by default in the Claude Code environment.

File reading and search: prioritize using Claude Code's native Read、Grep、Glob
Memory search: prioritize using nmem CLI, rather than abstract external runtime tools
Online supplementation: use the web tools available in the current session only when factual supplementation is truly necessary; do not treat going online as a default action
Path handling: directly use relative or absolute paths within the vault; do not rely on additional adapters
references files: read directly by repository path, for example references/profile-registry.md

Profile-driven version (multi-Profile) | Claude Code adaptation | Last updated: 2026-03-21

About this skill

Quick Usage Rules

Scenario Boundary Description

Style Source (Profile-Driven)

Profile Storage Structure

Recommended asset structure for each Profile

Migration Compatibility Notes

Initialization (required once the first time; can continue to add afterward)

Determine the Profile identity before initialization

Rules for distinguishing updates from new creation

Anti-patterns (must avoid)

V2 writing workflow (mandatory)

Phase -1: Create project structure

Phase 0: Build Brief

Phase 0.1: Human-only Brief Slots (anti-fabrication)

Phase 0.3: Profile Gate (choose first, then write)

Phase 0.5: Proactive context collection

Phase 0.6: Goldmine Budget Card (mandatory)

nmem memory search (mandatory)

Obsidian CLI search tips

Goldmine Pack (Soft-gated, enabled on demand)

Online research (triggered as needed)

Phase 0.7: Article Gravity Gate (pre-writing appeal assessment)

Phase 0.8: Reader Capture Audit (first-screen grip and content value check)

Gate A: Context sufficiency assessment

Phase 1: Socratic material supplementation (triggered as needed)

Phase 2: Mode routing

Phase 2.3: Article archetype routing (Profile-specific, on demand)

[Single-threading mode] Phase 3S: The main agent writes independently

Step 1: Pre-writing skeleton

Step 2: Full Writing

Step 3: Self-Check

Step 3.5：Human Feel & Flow Review

[Multi-threaded mode] Phase 3M: three sub-agents writing in parallel

Step 1: pre-writing skeleton

Step 2: three sub-agents writing in parallel

Step 3: Triple Lint

Step 4: Scoring, draft selection, controlled fusion

Phase 4: Final Draft Convergence

Publishing metadata convergence (triggered by platform gate)

baseline snapshot (for subsequent manual revision retrospective)

Decision and Fallback Mechanism

Resource Navigation (Progressive Disclosure)

Execution templates (ready to apply directly)

Template 1: Questions when Gate A is not passed

Template 2: Single-thread mode — structure selection output

Template 3: Multi-thread mode — main agent output after three drafts are completed

Self-iteration

Punchline generation (Profile version)

Reference resources

Claude Code execution notes